Cognirion.
39 (1991)
3
215-3s
Critical period effects on universal properties of
language: The status of subjacency in the acquisition of
a second language*
JACQUELINE
S. JOHNSON
University of Virginia
ELISSA L. NEWPORT
University of Rochester
Received
August
23.
1990.
final
revision
accepted
blay
16. 1991
Abstract
Johnson.
J.S..
guage: The
and Nevvport.
status of subjacency
E.L.. 1991. Critical period effects on universal properties of lanin the acquisition of a second language. Cognition. 39: 215-258.
Recent studies have shown clear evidence for critical period effects for both
first and second language acquisition on a broad range of learned, languagespecific grammatical properties. The present studies ask whether and to what
degree critical period effects can also be found for universal properties of
language considered to be innate. To address this issue, native Chinese speakers
who learned English as a second language were tested on the univ*ersalprinciple
subjacency as it applies to wh-question formation in English. Subjects arrived
in the U.S.A. between the ages of 4 and 38 years. They were immersed in
English for a number of years (a minimum of 5) and were adults at the time
of testing. Non-native performance on subjacency was found for subjects of all
ages of arrival. Performance declined continuously over age of arrival until
adulthood, (r = - .63). When immersion occurred as late as adulthood, performance dropped to levels slightly above chance. In all of the analyses performed, subjacency did not differ from language-specific structures in the degree or manner in which it was affected by maturation. These results suggest
that whatever the nature of the endowment that allows humans to learn lan*This research was supported in part by NIH Training Grant no. HD07205 to the University of Illinois
and by NIH Grant no. DC00167 to E. Newport and T. Supalla. We would like to thank Diane Lila-Martin
and Peter Cole for raising questions that helped stimulate this study. and Renee Baillargeon. Marie Banich.
Dedre Gentner. Gabriel Hermon. Doug Medin and Lydia White and three anonymous reviewers for their
helpful advice and comments. and Jeff Bettger for assistance in preparing the figures. This paper is based on
a portion of the Ph.D. dissertation of J.S. Johnson (University of Illinois. Urbana-Champaign.
1988. unpublished). Early portions of this work were presented at the 1987 Boston University Conference on Language
Development. Correspondence concerning this article should be addressed to Jacqueline Johnson. Psychology
Department. Gilmer Hall. University of Virginia. Charlottesv-ille. VA 22903. U.S.A.
OOlO-0277/91/$13.70
0 1991 - Elsevier
Science Publishers
B.V.
guage, it undergoes
ntature.
a \*er_ybroad deterioration
as learners become
increasingl)
Introduction
In most domains of learning, skill increases over development.
In contrast to
this pattern. however, recent research in language acquisition has shown that
the ability to aquire language does not increase with development:
rather, this
ability is at a peak very early in life and then gradually declines until maturation is complete. This reversed effect of maturation has been found in some
aspects of the learning behavior of other species. and is often referred to as
a crirical or semifive period for the acquisition of the behavior in question.
For language acquisition as well. we vvill use the term criricalperiod to refer
to the general phenomenon
of declining competence
over maturation.
By
using the term in this broad fashion we do not mean to imply vvhat mechanism
underlies this effect,’ but mean only to emphasize its maturational
underpinnings. Although a critical period was hypothesized
for language acquisition
as early as 1967 (Lenneberg,
1967). much of the empirical data remained only
suggestive until recently. The empirical studies now show clear evidence for
critical period effects in language acquisition, both in the acquisition of the
grammar of a first language (Curtiss. 1977, 1988; Newport & Supalla, 1990)
and of the grammar of a second language (Johnson & Newport, 1989; Oyama,
1978; Patkowski, 1980). This paper concerns itself with extending our understanding of the breadth or limitations of such effects.
Several studies testing the critical period hypothesis have evaluated and
compared the grammatical capabilities of adult and child language learners
after they had been exposed to the target language for many years.’ All of
the studies taken together have shown an effect of age of acquisition on a
‘That is. we will initiallv ask the question without prejudging what the degree or quality of such maturational change may be (e.g..‘is it a sharp qualitative change vs. a gradual quantitative one?) and what the nature
of the underlying maturational mechanism may be (e.g.. is it a change in a special lanyage faculty vs. a more
general change in cognitive abilities?). We wll. hou-e\er. return to both of these issues in the General
discussion.
‘There are also many studies which aim to test the critical period hypothesis. but which do so by examining
children‘s and adults’ abilities in a language within the first hours or months of exposure to the language. As
Krashen. Scarcella. and Long (1982) and Johnson and Newport (1989) have discussed. this type of study
addresses a quite different issue. namely the effect of age of exposure on the rare of earlv learning, rather
than the effect of aSe of exposure on ultimate competence in the language. Interestingly. the’initial rate studies
tend to show advantages for adult learners. uhile the ultimate competence studies (revieued above) shorn
advantages for child learners (Krashen et al.. 1982). The present study focuses on ultimate competence.
However. a full account of the critical period for language acquisition must explain both of these phenomena.
broad range of grammatical constructions.
tested in a variety of ways. In one
of our own studies - a study which will be referred to frequently in this text
- native Chinese and Korean speakers who had learned English as a second
language were tested on 12 basic properties of English grammar. While these
properties posed varying degrees of difficulty for late learners of English, all
of the structures showed critical period effects (Johnson & Newport. 1989).
It is clear that the critical period has an impact on the acquisition of many
parts of grammar.
This apparent breadth of critical period effects might lead one to prematurely conclude that whatever capacity underlies language acquisition. it undergoes some decav in its entirety over maturation. This conclusion is premature only because virtually all of the studies to date have tested parts of the
grammar that are considered specific properties of the target language. and
which to many linguists are mere details of the grammar. Properties which
are more general to all languages of the world, and which are hypothesized
to form part of the innate knowledge the child brings to the learnmg process,
may or may not show an effect of maturation.
These more general. or universal, properties are viewed as aiding the learning process by constraining
the types of hypotheses one might make about the language. In contrast,
language-specific
structures such as those tested in previous critical period
experiments must be learned from exposure to the details of a particular input
language: their form is not innate. The question of whether universal constraints show effects of age of acquisition is thus a separate question from
whether the learning of language-specific
structures shows such effects. Various patterns of outcomes on these two questions may lead to different conceptions of the nature of maturational change in language acquisition. One possibility is that, while there is a decline over age in the ability to learn languagespecific structures as shown by previous studies, the constraints underlying
the language universals might be maintained throughout life. An alternative
possibility is that there is a decline over age in both the learning of languagespecific structures and the use of language universal constraints. This latter
outcome
would suggest a much more general maturational
change in a
mechanism broadly affecting all aspects of language acquisition.
To make this distinction clearer, an overview of the concept of universal
grammar will first be presented. Following this. we will return to the issue of
a critical period for language and in so doing will discuss in more detail what
aspects of grammar we already know are affected by maturation in second
language acquisition.
Finally, the current state of universals and their relationship to second language acquisition will be reviewed briefly.
Unir*ersal
grmmnnr
and lenrnnbili!\
Universal grammar refers to properties of grammar which are common to all
languages of the bvorld. For the last two decades, following Chomsky (1965.
lYSl), it has been the goal of many linguists to discover what these universals
are. The interest in these language universals, for linguists and psychologists
alike. stems from the belief that these universals constitute the innate knowledge needed for learning language. These unive_rsals. along with some
parametric variation, are thought to arise from the fact that there are constraints as to what can constitute a human language. They make learning
language possible by limiting the number and type of implicit hypotheses a
person forms when acquiring a language.
While the psychological status of many formal linguistic principles remains
open, few scholars question the premise that some constraints must exist for
language learning to be possible. Without expectations
or biases as to what
form a language can take, language can be shown to be unlearnable.
It can
be shown that, for every set of utterances directed to a child. there is an
infinite number of hypotheses or grammars which could account for that input
(Gold, 1967; Wexler & Culicover. 1980). The only way of making an infinite
number of possibilities finite is by constraining
them: the child must not
consider all of the possible hypotheses,
but only those that are natural (i.e..
within the realm of possibilities for a human language). Determining what is
universal or systematic about languages of the world is one way of discovering
what kind of constraints and biases the child might have which make learning
possible.
Some universal properties are thought to be entirely universal in languages
of the world, while other universal properties
involve a small degree of
parametric variation. In this latter case, while languages are universally very
similar on the property in question, they are not the same in detail; rather,
there is a highly restricted range of variation on some particular aspect of the
known as a parameter.
The possible values of this
universal property.
parameter found in languages of the world are called parameter settings. Note
that the principle still applies universally across languages of the world, yet
because the principle operates parametrically,
slight variation will exist. In
such a case, the goal of delimiting the number and type of hypotheses that
the child makes is still satisfied as long as the parameters are finite in number
and possible values, and as long as there is some principled way for the child
to choose among these values.
In contrast to universal properties of grammar, language-specific
structures
are those properties of languages which vary much more widely. Variations
on language-specific
properties are more detailed and idiosyncratic to a particular language. Operationally,
they are structures for which investigators
tinivtwal propertirs
1IY
have not been able to find a principled description cross-linguistically,
and
for which such a principled description does not look possible. They include
(but are not limited to) such properties as lexical subcategorization,
certain
aspects of question formation,
morphology
and the presence or absence of
other closed class items. Again, they constitute details about a language which
must simply be learned. For example. despite the semantic similarity of the
words speaks and spays, a person learning English would have to discover that
says requires an object while speaks does not. As another example, one must
learn when a past tense ending is required on the verb, and moreover. learn
what form of ending that verb requires. While it is possible that some constraints governing language-specific
properties will be found upon future inquiry. it is obvious that these constraints would nevertheless
have to allow
for substantial variation. The boundaries set by these constraints would be
very wide. and within these boundaries the details of these properties would
range freelv. In contrast, the boundaries set by universals are narrow. and
while allowmg some minimal variation, do so in a highly constrained manner.
Turning now to the second language research, the types of language structures that have previously been shown to be affected by maturation will be
reviewed. These structures are considered to be language specific and hence
should give the reader more familiarity with what type of structures fall within
this category.
Critical period effects in second
strrrctwes of English
Iflngunge acquisition
for lflngiiage-specific
Three studies have reported critical period effects for the acquisition of grammar of a second language (Johnson & Newport, 1989; Oyama, 1978: Patkowski, 1980). Two of these (Oyama, 1978; Patkowski, 1980) examined the
effects of age of exposure to English as a second language using various global
measures of proficiency in the language. The third is a previous study of our
own (Johnson & Newport, 1989). which examined subjects’ knowledge of
particular grammatical structures in the second language (English). It also
addressed a number of questions concerning the rate and nature of the critical
period effect that had not been addressed in these earlier studies. Because
of this, this study will be used as a yardstick for interpreting
the results of
the present studies and will therefore be reviewed here in detail.
In Johnson and Newport (1989), henceforth J & N, the primary goal of
the research was to determine whether critical period effects for second language acquisition existed. If there were a critical period for second language
acquisition, then learners beginning their exposure to the language in early
childhood should reach higher levels of proficiency
in that language than
those beginning at later ages, particularly adulthood. We tested this predic-
__
“0
J.S. Johnson
and E. L. .Vr~c~port
tion by comparing the grammatical proficiency in English attained by native
Korean or Chinese speakers who had arrived in the U.S.A. between the ages
of 3 and 39, and who had lived in the U.S.A. for an average of 10 years by
the time of testing.
These subjects were tested on 12 basic structures of English grammar with
an aurally presented
grammaticality
judgment task. The structures tested
include past tense, plural, third person singular. present progressive, determiners, pronominalization,
particle movement, subcategorization.
auxiliaries,
yes/no questions, wh-questions and basic word order. While these 12 structures operate within universal constraints at some level. the. violations we
formed were within the set of acceptable
variations found among human
languages of the world; the particular aspects of the rules we tested were
consequently
language-specific
features of English.3
Both correlational
and t test analyses demonstrated
a clear and strong
advantage for younger arrivals over older arrivals. Figure 1 shows overall
performance
on the test, grouped and plotted by age of arrival in the U.S.A.
This figure represents the mean score correct (out of 276 possible) for each
of the learning groups. As this figure shows, there is a systematic relationship
between performance
and age of arrival in the U.S.A. The 3-7-year age of
Figure
1.
The relationship benreen age of arrival in the U.S. A. and total score correct
on a test of English grammar (from Johnson & Nebc,port. 1989). Reprinted
with permission from Academic Press.
270 _
.
260 .
260 240 230 220 210 200
Native
3-7
8-la
Age
of
11-15
17-39
Arrival
-There were a few exceptions to this: some sentences testing pronominalization violated binding theory.
Most of the test items testing this rule type. however. varied gender agreement and number agreement. not
distance between the pronoun and its referent noun.
arrival group performed on this test at the native level. In each of the remaining groups, as age of arrival increased. performance
became significantly
poorer.
The degree to which the rule types were mastered by the older learners
varied markedly. but a significant correlation
between age of arrival and
performance
appeared on all of the 12 rules types tested. In all cases adult
arrivals performed more poorly than younger aged arrivals. and in 2 of the
12 rules they performed below chance.
To summarize. this research showed that on basic properties of English
grammar there is a systematic relationship
between age of exposure and
performance
many years later, and thus there is an apparent effect of maturational state on the learning of a second language. More specifically, this study
showed a decline over age in second language learners’ control over some of
the most basic, language-specific
aspects of English. The studies in the present paper are designed to test whether age of acquisition also influences the
second language learners’ use in English of principles of syntax thought to be
universal.
Second lnngrrage acquisition
nnd adldt nccess to universd
grnrnmar
There is a growing literature on the relationship of universal grarnmar (UG)
to second language acquisition. With increasing scholarship on the principles
of UG, there has also been an increasing number of theoretically
sophisticated papers which have related UG to second language acquisition. Because
UG is hypothesized to form the innate knowledge underlying language. many
theories have rested on the unnecessary assumption that the adult learner
must therefore have full access to this knowledge (Flynn, 1983, 1984; Liceras,
1985; Mazurkewich,
1984; Phinney.
1987; Ritchie, 1978; White, 1983a,
1984b) .’
In singular opposition, Clahsen and Muysken (1986) have argued that the
adult second language learner does not have access to UG. More specifically,
these investigators have claimed, based on a comparative analysis of children
and adults learning German word order, that adults are not guided in their
‘Some researchers working under this assumption have predicted that UG. while available. will not work
in the same way as it does in first language acquisition. They su_egest instead that the parameter settings of
the first language will have some effect on the ultimate success and accurate application of universal principles
to the second language. Unfortunately. some of this work conflates child language learning with first language
acquisition. and adult language learning with second language acquisition, making it difficult to understand
whether in their view the influence of a first language’s parameter setting on the second is a result of having
already set that parameter in acquiring a first language. or rather. is a result of having matured with a particular
first language parameter. The latter of course would be a critical period effect. whereas the former would be
predicted to exist among second language learners of all ages. whether younger or older.
learning by the universal principles governing word order. In rebuttal, duPlessis and colleagues have proposed that the errors of the adults in this studt
reflected a difficulty with setting or resetting the parameter of the universal.
and not with obeying the universal in question. claiming that all of the errors
made by the adults during the process of language learning were within the
set of possible grammars permitted bc the universal (duplessis, Solin. Travis.
& White, 1987). Whether or not in this case the adults’ errors conformed to
principles of UG, duPlessis et al.‘s general point is legitimate: evaluations of
the accessibility of UG must distinguish between the ability to obey a principle in general and the ability to learn the specific parameter setting of the
second language.
Only very recently has any additional work asserted that adult learners do
not have full access to UG (Bley-Vroman
& Chaudron,
1987: Bley-Vroman.
Felix, & Ioup, 1988: Schachter,
19S9). Two studies (Bley-Vroman
et al..
19SS; Schachter,
1989). conducted
simultaneouslv
with the present work.
claim to support the notion that UG is not working at full capacity in the
adult second language learner. These studies will be described. first to show
that there is some suggestive evidence indicating that UG may not be working
at full capacity in the adult learner, and second to show at the same time that
stronger empirical evidence is nevertheless
needed to support this important
claim.
In both studies, post-pubescent
second language learners of English were
evaluated
in their competence
on the universal principle subjacency as it
applies to English. Briefly, the subjacency principle prevents wh-question
formation from occuring in certain types of complex sentences (see the next
section for a more detailed description).
Bley-Vroman
et al. (1988) tested
native Korean speakers’ observance of subjacency (and other structures) in
English with a written grammaticality
judgment test. The results showed
below-native
performance
on subjacency. suggesting that adult learners do
not have full access to UG. The single limitation of this study is one of scope:
the study utilized a small and highly limited set of test items (only S items
were used to test subjacency).
In a more extensive study, Schachter (1989) tested three groups of subjects
who differed in their native languages: Indonesian. Chinese and Korean. The
test was a written grammaticality
judgment task which involved 2-t ungrammatical questions involving subjacency violations. and 24 grammatical declarative sentences with the same complex structures that were used to test
the subjacency principle. The declarative sentences were intended to provide
a means of determining whether subjects knew the structures on which subjacency would be tested well enough to permit evaluating their results on subjacency: these sentences constituted
the syntacric rest. The ungrammatical
questions formed the subjacency test. The data were analyzed dichotomously:
subjects either passed or failed the syntactic and the subjacency test according
to set criteria.
No differences were found between the three different language groups.
Most of the subjects passed the syntactic test but failed the subjacency test.
suggesting again that adult learners of a second language do not have full
access to UG (specifically. to the subjacency principle). There are. however.
some limitations in the design of the test and the manner of evaluation which
allow for alternative accounts of the results.
First. the test was not designed to rule out a simple response bias. Passing
the syntactic test and failing the subjacency test both require a yes response.
To obtain these results, then, subjects only needed to respond yes to all of
the test items. independent
of the structure being tested at that time. and
similarly independent
of any real knowledge of English syntax.
The test was also not designed to provide clear evidence that a subject is
obeying subjacency. The subjacency principle prevents wh-question formation
from occurring only in certain types of complex sentences; in other complex
sentences question formation is grammatical.
Since subjects were tested on
wh-questions
that violate subjacency and not on wh-questions
that observe
the subjacency principle. rejection of these sentences could have happened if
subjects believed that question formation in any form is ungrammatical,
or
more reasonably. that question formation in any complex sentence is ungrammatical. Subjects’ ability to distinguish question formation which violates subjacency from question formation which does not violate subjacency was not
tested.
In addition, since Schachter presented only the results of a dichotomous
scoring procedure (i.e.. how many subjects passed or failed each test by her
criteria), one cannot tell the degree to which subjects erred on subjacency.
Previous literature on adult language learning has consistently shown graded
effects on grammatical competence.
Dichotomous scoring eliminates the possibility of seeing similar probabilistic tendencies in the observance of subjacency, as compared with absolute adherence to the principle.
In sum, the Schachter (1989) and Bley-Vroman
et. al. (1988) studies provide provocative
evidence that UG may not be fully available to the adult
language learner. However, the work needs to be replicated and expanded
upon to determine whether this is true.
In the current studies, also testing the subjacency principle, we hope to
provide a more determinate
test of the availability of UG to the second
language learner by eliminating some of the design problems mentioned
above. Also. using a continuous rather than a dichotomous
evaluation of
competency,
we intend to provide a more complete understanding
of the
second language learners’ knowledge, or lack of knowledge, of subjacency.
The mbjacency
principle
As in the studies discussed above, a test was constructed
using structures
which are governed by the universal syntactic principle, subjucetzcy. Historically, this universal
principle
was first called the A-o\*er-A principle
(Chomsky. 1964, 196S), then island consrrclinrs (Ross. 1967), and only more
recently has it been reanalyzed,
and renamed subjacency (Chomsky, 1981).
Subjacency was chosen for study because. of the universal principles that have
been proposed, this one is relatively well understood and has stood the test
of time. For present purposes, it is less important how the current formulation
of the subjacency principle works; what is important is that, regardless of the
particular linguistic theory one adopts, the descriptive facts have remained
the same. In addition, the psychological validity of subjacency is supported
by empirical work with first language learners of English; this work has
suggested that children observe subjacency as soon as they acquire the relevant structures
to which it applies (Otsu. 1981). The present study asks
whether this is also true of learners exposed to English as a second language,
and whether it is true for younger and older learners alike.
The principle
As stated earlier, the subjacency principle places restrictions on the types of
structures from which wh-movement
can take place. In general. wh-movement can occur from very far down in a sentence. This can be seen in (la)
and (lb) below. Despite the multiple embeddings in the declarative sentence
(la), a grammatical question can be formed by extracting a wh-item out of
the deepest embedding, shown in (lb):
(la) John heard that Mary said that Tom loves Susie.
(lb) Who did John hear that Mary said that Tom loves?
There are certain types of embeddings.
however, out of which wh-items
cannot be moved. Figure 2 presents several sentences. along with their tree
diagrams, to illustrate the subjacency principle. Sentence (2a) is a declarative
sentence from which we can form the grammatical question shown in (2b).
Sentence (2~) is a similar declarative sentence, yet the comparable question
in (2d) is ungrammatical.
The subjacency principle handles this phenomenon
as follows: it states that movement of a word from its deep structure position
in a sentence may not occur over more than one bowding
node within a
cycle. In English the bounding nodes are the categories S and NP. The first
question (2b) is grammatical because, when the wh-word is moved cyclically
(by camp-hopping),
it passes over at most one bounding node in any one
cycle (circled). The second question (2d) is ungrammatical
because the wh-
L’nir~ersal properties
Figure 2.
Sentence trees illttstrnting grammatical
T\
camp
nnd tmgrclmmc~tical bth-movement.
A
My mother
NP
TP\
heard
7\
wmp
that
A
Mrs.z”es
V
is buying
/
vp\
NP
a computer/what
2a. My mother heard that Mrs. Jones is buying a computer
2b. What did my mother hear Mat Mrs. Jones is buying?
7\
camp
A
My mT-ter
heard”
/
“‘\
/T
theNnks
/
ctEp
“\
,@,
Mrs. Enes
/ “‘\
V
NP
is buying
a computer/what
2~. My mother heard the news that Mrs. Jones is buying a computer.
26. ‘What did my mother hear the news that Mrs. Jones is buying?
word passes over two bounding nodes in the second cycle.
Sentences like (2d) and other extractions from complex NPs are universally
ungrammatical.
However, to a limited degree, languages vary on which other
extractions
violate subjacency.
In more formal terms, what constitutes
a
bounding node forms a parameter upon which languages may vary. The set
of possible bounding nodes is limited to three or four possibilities:
the
categories S’, S and NP (and possibly PP). At the risk of oversimplifying,
there are thus three possible parameter settings, formed by pairing all possible combinations
of bounding nodes (S’ & S, S & NP, S’ & NP). Any whmovement which requires movement over S’, S & NP (as happens in extrac-
tion from complex NPs) will be ungrammatical
in every language. regardless
of the particular choice of bounding nodes in that language. Thus many
subjacency violations in one language are also ungrammatical
in languages
universally.” These we will refer to as irri~rirrnf subjacency violations. However. certain wh-movements
(e.g.. movement over S, S’. and another S, as
happens in extraction from WH-complements)
will be ungrammatical in some
languages but grammatical in others. These we will refer to as parcunetricalfq
rwying subjacency violations.
Because of the manner in which the subjacency principle works. it is possible
to create a test which asks whether learners observe the invariant aspects of
subjacency in English. as well as whether they observe the parametrically
varying aspects of subjacency in accord with the parameter settings of English. With such a test. effects of age of acquisition on subjacency can be
examined in two ways. First. we can ask whether age of acquisition is related
to performance
on subjacency overall. pooling items that test the two aspects
of subjacency. If adult learners perform as well as natives on this pooled set
of items (and they perform in appropriate
ways on items controlling for other
grammatical factors), we can conclude that age of acquisition does not affect
the use in a second language of this language universal, either in its invariant
or its parametrically
varying aspects. On the other hand. if adult learners do
not perform as well as natives on this pooled set of items. we can examine
their performance
in more detail to determine
whether age of acquisition
affects only the parametrically
varying aspects of subjacency, or whether it
affects the invariant aspects of subjacency as well. Poor performance
on the
invariant aspects of subjacency would suggest that maturation may lead learners to violate language universals, entertaining
hypotheses
about English
which are thought to be outside the possible class of human grammars. Good
performance
on the invariant aspects but poor performance
on the parametrically varying aspects, however, would suggest that maturation does not lead
learners to violate language universals in the extreme. but only to have difficulty
determining the parameter settings of a new language.
The primary goal of this paper is to determine whether there is or is not
a critical period effect for the application of universal properties to second
language learning. To meet this end, two studies are presented. Both studies
‘Some work has shown that certain dialects of Swedish violarc all known parameter values for subjacenq
(Atwood. IY76). If future work does not find a parameter value which Swedish obeys. then obviously it would
be incorrect to say that there is an invariant boundar)- for subjacency (a boundary which separates languages
from non-languages). If the work on Swedish holds up. then violations of this “invariant” boundary by adult
learners must be interpreted as their tendency to form rare hypotheses about language. rather than as their
tendency to form wr~~orwc~/hypotheses about language.
Utlir~erstrlproprrtirs
227
involve the use of subjects who are native speakers of Chinese who have
learned English as a second language. Chinese was considered to be an ideal
choice as a native language for two reasons. One, it allows for direct comparisons between the results of this study and the results of J & N. where we also
used native Chinese (and native Korean) speakers. Secondly. wh-questions
in Chinese do not involve movement. Therefore,
although Chinese does not
violate subjacency. subjacency does not apply to wh-question formation in
Chinese.
(There is evidence
suggesting that subjacency
might apply in
Chinese at the level of logical form for wh-questions,
and in the syntax for
topicalization.
While these issues are unresolved,
it is clear that subjacency
does not apply to wh-question formation in Chinese.) What is important here
is that subjects cannot get the correct answers for English subjacency structures by strict transfer from their native language (i.e., by superficial translation and borrowing).
The first study evaluates the availability of subjacency in native Chinese
speakers who were first immersed in English as adults. The second study
evaluates the availability of subjacency in native Chinese speakers who were
children when they first learned English. Performance
is evaluated over different ages of learning and consequently
provides detailed information about
the relationship between age of learning and the use of language universals.
The two studies together ask whether and in what way subjacency is affected
by the maturational
status of the learner.
Study 1
As noted, adult learners of English have been shown to have lasting deficiencies on a wide range of basic properties of English grammar. This study asks
whether or not similar deficiencies exist for universal properties.
The questions specifically addressed in this study are as follows:
(1)
(2)
(3)
In learning wh-question formation in English, do native adult Chinese
speakers obey subjacency even though they have never seen it apply in
their first language?
If performance
on subjacency is less than native, is performance
on
subjacency
nevertheless
high relative to language-specific
structures
(thus granting subjacency a privileged status with regard to maturation),
or is performance
on subjacency affected by the adult learner’s maturational state in a manner similar to language-specific
structures?
If performance
on subjacency is deficient, how deficient is it? To what
degree, if at all, is subjacency available to the adult learner?
(4)
(5)
Does performance
vary depending upon what types of sentence structures subjacency is tested in? If performance
does vary. does it depend
upon whether the subjacency violation is an invariant or a parametrically
varying one?
Finally, does the pattern of the adult learners’ performance
indicate that
they are adopting a parameter value other than the one for English?
Method
Subjects
Subjects were 23 native Chinese speakers who learned English as a second
language as adults. To qualify as an adult learner of English. subjects’ arrival
in the United States, and consequently
their first immersion in the English
language, had to occur after the age of 17. The 23 subjects meeting this
criterion arrived in the U.S.A. between the ages of 15 and 3s. with a mean
age of arrival of 25.8.
Although subjects were not immersed in English until after the age of 17,
all but one of the subjects had at least some English training through formal
classroom instruction at an earlier age. The majority of the subjects began
formal English training when they began high school. at an average age of
13.5 (range = 9-14). They generally continued English classroom training
throughout
high school and most of college, with a mean of 6.8 years of
instruction and a range of 4-12 years. This earlier and extensive training in
English (at least in terms of number of years of exposure) naturally brings
up the question as to whether these subjects could be truly classified as adult
learners of English. Our previous research suggests. however. that whatever
natural endowment
youth confers upon the language-learning
process, it is
limited to the natural learning situation (i.e., immersion) and not to the highly
formal classroom setting in which our subjects were exposed to English prior
to arriving in the U.S.A. (Johnson & Newport, 1989); in this prior study, age
of beginning classroom instruction did not correlate with ultimate proficiency
in English, while age of immersion in English in the U.S.A. did. For this
reason, then, age of exposure to English in the present study is calculated on
the basis of age of immersion.
To qualify for this experiment
it was also necessary for subjects to have
had sufficient experience
with English to be considered
at their ultimate
attainment in the language. Previous research has shown no effects of length
of experience with the language for adult learners of a second language after
5 years of immersion in the target language (Oyama, 1978: Johnson & Newport, 1989). Subjects in this study, consequently,
were required to have a
Lrrfibwsal proprrtics
2’9
minimum stay in the U.S.A. of 5 consecutive years before the time of test.
The resulting group had an average stay in the U.S.A. of 6.2 years, with a
range of 5-12 years. Additionally,
to ensure ample exposure to English. and
to ensure some homogeneity
of social background,
subjects were selected
from the graduate student, post-doctoral,
and faculty population at an American university (University of Illinois). Subjects were recruited through letters
and by word of mouth.
The purpose of the study also requires that subjects have sufficient competence in English to be tested on a principle which involves the use of complex
sentences. Our best hope of satisfying this requirement
was to select subjects
who could be considered among the better speakers of English, as well as to
eliminate subjects whose competence
in English was so poor that they could
not deal adequately with the sentence structures to which subjacency would
be applied. Any subject who did not perform above chance on the simplest
structure of our test, wh-movement
in simple sentences, was eliminated. One
subject was eliminated for this reason.
The resulting 23 subjects consisted of 14 males and 9 females. who at the
time of test were between the ages of 23 and 4, with a mean age of 32. Their
language environment
at time of test most often was bilingual: speaking their
native language in the home, and English at school and work. For some of
the subjects, however, particularly those that were unmarried, the language
environment
was essentially all English. Previous research shows no differences in performance
due to percentage
of day-to-day
exposure with the
second language, at least with subjects who had been exposed to the second
language for a number of years and had substantial day-to-day
exposure
throughout those years, as is the case here.
To provide a baseline performance
on the subjacency test, 11 native speakers of American English were also run. All were undergraduate
students
attending an American university (University of Illinois) and were recruited
through posted sign-up sheets.
Procedure
The subjects were tested on their knowledge of subjacency and related
English structures through a grammaticality
judgment task (see Materials).
Subjects listened to aurally presented English sentences which were either
grammatical or ungrammatical,
and were asked to make a judgment about
their grammaticality.
The test sentences were recorded on tape in a native
American female voice. Each sentence was read twice, with a 1-2-second
pause separating the repetitions. They were said clearly, with normal intonation, at a slow to moderate speed. The ungrammatical
sentences were spoken
with the intonation
pattern of the grammatical
counterpart.
There was a
6-9-second
delay between the different sentences.
This was an adequate
amount of time for the subjects to make a decision.
Subjects were tested individually in the laboratory.
They were told that
they would be hearing a series of tape-recorded
sentences, some of which
were ungrammatical
or ill-formed in some way. and some of which were
perfectly fine, grammatical sentences in English. They were then instructed
to listen to each sentence carefully and make a judgment as to whether or
not the sentence is a grammatical
sentence in English, circling yes if they
thought the sentence was fine and no if they thought it contained an error of
some sort. They were instructed to respond to all of the sentences, guessing
if they were unsure. It was made clear to the subject that if the sentence was
incomplete or otherwise incorrect they should regard it as ungrammatical.
There was one exception to this rule: subjects were explicitly told to ignore
the who/whom distinction (sometimes used in English to mark nominative
versus accusative case). The experimenter
said to the subject that they were
probably told that it is proper to say “whom did the boy kiss?” instead of
“who did the boy kiss?“, but that for this test they were to ignore this distinction since colloquial American English tends not to use it. It was made clear
that they were not to mark any sentence wrong solely on the basis that they
thought it should have had a “whom“ instead of a “who”. To avoid giving
clues to the subject, the experimenter
did not face the subject during the
testing session while the tape was going. Subjects were given a break halfway
through the test, but were told prior to starting that they should tell the
experimenter
to stop the tape at any time if they needed to break sooner,
either if the tape was too fast for them or if they were simply getting tired.
Following the grammaticality
judgment
test, information
was gathered
about the subject’s age, occupation,
education and language background.
including age and years of English training prior to coming to the U.S.A.,
and current exposure to English. A subset of the subjects. in addition, took
a short written comprehension
test to test whether or not they were correctly
interpreting
the questions with subjacency violations. Since this test forms a
very small part of this study. details of it will be described and presented in
the Results section.
Materials
A grammaticality
judgment test with a total of 180 test sentences was
devised to test the subjacency principle. The test was formed around three
different types of sentence structures to which subjacency is relevant in English: noun phrase complements
(NP-camp), relative clauses (RC) and WHcomplements
(WH-camp).
Twelve basic sentences were created for each of
the three sentence structures (NP-camp, RC and WH-camp),
making a total
L’tti~~ersalproperties
‘31
of 36 basic sentences. These 36 sentences were each presented in four different forms. as described below. An example of each of these four forms for
each of the three sentence structures can be found in Table 1 as well.
This sentence type is the declarative form of the sentence:
A. Declarative.
12 grammatical declarative sentences of each of the three sentence structures
(NP-camp, RC and WH-camp) were created (see la. 2a and 3a in Table 1).
The declarative form of these structures was included in this test to make
sure that subjects found these structures and the particular sentences used to
instantiate these structures as grammatical,
since some form of them would
be used to test the subjacency principle.
This second sentence ty:pe is a wh-question
B. Subjacency
violation.
which violates the subjacency principle; it consequently is ungrammatical (see
lb, 2b, and 3b in Table 1). This violation is instantiated
in the simplest
structure possible needed to test subjacency; no adjectives, adverbs or unnecessary embeddings are included. In this sentence type subjacency is violated by moving a wh-word over two bounding nodes. It forms the test case
of the subjacency principle, for subjects who accept sentences of this form as
grammatical are failing to obey the subjacency principle.
C. Control.
A third type of sentence, a control sentence, was created by
moving a wh-word in a similar sentence but over only one bounding node
(see lc, 2c and 3c in Table 1). This sentence structure is grammatical and
parallels closely the structure involving the subjacency violations. It acts as a
comparison sentence designed to test whether subjects will allow wh-items to
be moved out of complex sentences in general. If, for example, subjects
correctly reject the subjacency violations, but also incorrectly reject the control sentences, then they could not be credited with obeying subjacency for
they fail to allow movement to occur within any complex sentence.
D. No subject-auxiliary
inversion.
The fourth sentence type contains errors of subject-auxiliary
inversion (see Id, 2d and 3d in Table 1). It involves
what would be a grammatical wh-question,
as in the control sentence, but
forms an ungrammatical
sentence by failing to invert the subject and auxiliary. This type of sentence allows comparison of subjacency results to results
on language specific structures. Both apply to wh-movement;
however, one
is particular to English (inverting the subject and auxiliary) and one is universal (subjacency). If subjacency has a privileged status with regard to maturational effects, then late learners should perform better on the subjacency
violations than they do on subject-auxiliary
inversion violations.
Table 1.
Examples of tire folrr senteme
(/VP-camp. RC, WH-camp)
(la)
t>‘pes for
ench of the three serlteme
The teacher knew the fact that Janst liked math. (dcclar.)
(lb)
“What did the rexher
( Ic)
(Id)
What dud the teacher knon that Janet liked’? (control)
know the fact that Janet liked? (wbj.)
‘What the teacher did knou that Jsnet liked? (no inv.)
SS rrloriw
(13)
chrst3
The policeman who found Cath! >hould get a reward. (de&w.)
(2b)
*Who should the policeman u ho found get a reward? (wbj.)
(3)
What should the policeman &ho found Cathy set’? (control)
(2d)
“What the policeman who found Cathy should get?
WH-complernorr
(33)
structures
(no inv.)
sm,crurt~
Sally natched how Mrs. Gomez makes her cookies. (de&r.)
(3b)
*M’hat did Sally watch how Mrs. Gomez makes? (subj.)
(3)
Who did Sally show how .Mrs. Gomsz maker her cookie\? (control)
(3d)
“Who Sally did show how Mr$. Gomez make\ her cookie\‘? (no in\.)
In addition to these sentences, 12 simple grammatical wh-questions were
also included in the test. It was deemed that if subjects did not perform above
chance on these simple sentences, they did not have enough competence
in
English to be evaluated on more complex structures. As described above (see
Subjects), one subject’s results were eliminated from the data set because he
did not perform above chance on these structures.
The 36 sentences that were presented in the four different forms, plus the
12 simple sentences, constitute the 156 core test items. In addition to these
items there were 24 filler sentences. These fillers were complex sentences
with an equal number of grammatical and ungrammatical
exemplars and only
served to increase the variability of the sentence types; they are not included
in any of the analyses. Thus, although 180 test items were responded to, only
the core of the test - are included in the
156 items - those constituting
analyses and figures which follow. Of these 156 items, 72 are ungrammatical
and 84 (72 + 12) are grammatical.
To reduce the possibility that subjects’ difficulty with a structure was due
to vocabulary rather than grammar, the sentences were constructed to contain
only relatively high-frequency
words. most of which were only one or two
syllables in length.
The test was divided into two halves. An equal number of exemplars of
each of the three structures and each of the four sentence types plus the
simple questions were presented in each half. Only one grammatical and one
L’f1i~w.d proprrrirs
ungrammatical
version of the same sentence was presented in each half of
the test. Within each half of the test the sentences were randomized in such
a way as to prevent clustering of similar test items or similar sentences. and
prevent long runs of Fes or no responses.
Responses to the sentences with the subjacency violations are obviously
crucial for determining whether subjacency is preserved in adult learners. The
responses to the other test forms, however. are equally important. for they
help us interpret the results on the subjacencv violations. That is. Lvhether
subjects are responding primarily correctly or-incorrectly
on the subjacency
items alone will not tell us whether subjacency is preserved in adult learners,
for alternative explanations for the results would remain depending upon the
outcome of the results on the other sentence types tested. More specifically.
performance
on these sentence types allows us to determine whether subjects
know wh-movement
and complex sentence formation
we!! enough to be
evaluated on subjacency. whether they adopt any response bias in their performances, and whether they distinguish between grammatical wh-movement
in complex sentences and ungrammatical
wh-movement
in complex sentences
which violate subjacency. The pattern of results across a!! of the sentence
types. and not the results from any one sentence type, will determine what
conclusions can be made about subjects’ knowledge of subjacency.
As noted in the Introduction
of this paper. it is important to test not only
the subjects’ ability to obey a universal constraint as it applies to the second
language, but also to test their ability on those structures within this constraint
which are and are not parametrically
varying. The test was designed to discriminate these aspects of the principle. Thus, if subjects do not obey subjacency as it applies to English, the test allows us to determine whether or not
subjects are at minimum obeying subjacency on sentence structures which are
invariant across human languages. If subjects obey subjacency on structures
which are invariant, this would suggest that UG is fully available to them and
that their only difficulty lies in their ability to correctly set the parameter. If,
on the other hand, subjects violate the constructions
which are universally
invariant, assuming theory to be correct. then subjects are willing to form a
grammar which is outside the possible class of human grammars.
The test was also designed to determine
whether or not subjects obey
subjacency
by adopting a different
parameter
value than English. The
parameter values form subsets of permissible and non-permissible
structures.
Over the control sentences and the subjacency violation sentences. for each
of the three structure types. three parameter values are represented:
they are
summarized in Table 2. Two of the test sentence types are allowable in any
language with movement.
They are the control sentence for the relative
clause. and the control sentence for the WH-complement.
In both of these
Table 2.
Subjacetq
status of the test Sentences nccording to three pnrnmeter rxltres
L’iolutrs subjacency with the follouing hounding node\:
-ret 5entence
Evampls
S’SNP
e.g. (Russian)
RC control
(Zc)
WH control
(.;C)
NP control
WH subj. \iol.
(Ic)
(3)
NPsubj. \iol.
(lb)
RC subj. viol.
(Zb)
SNP
S’ NP
(Englkh)
(It;lllan)
< <.r**
<.<F.‘*
C:L*Sr
.<l’<.<_&
<T.Lc*
-r* +-r
<:<<.$*
d‘I-6’
*air-i;r
cases only one possible bounding node is passed, an S. There is one test
sentence type which is allowable in any language except the most restrictive.
This sentence type is the control sentence for the NP-camp. which is a verb
phrase complement
(VP-camp).
Wh-movement
in a VP-camp passes over,
in one of its cycles, an S and S’. Any language which has both S and S’ as
bounding nodes should find this sentence ungrammatical
(for example. Russian, whose bounding nodes are hypothesized
to be S, and S’, (Erteshik.
1973)). The three types of structures that are called subjacency violations in
this test are of course ungrammatical to languages like English. whose bounding nodes are S and NP (German is another example). One of those sentences
is grammatical
in a language like Italian or Spanish (or it seems, Romance
languages generally) (Rizzi, 1978; Sportiche, 1981; Torrego, 19S-l). This sentence is the WH-camp subjacency violation. Because the complementizer
is
filled with a wh-word, wh-movement
from the downstairs S has to pass two
Ss and one S’. In most Romance languages the bounding nodes are S’ and
NP, therefore in this sentence only one bounding node is passed. The other
two sentence types which constitute subjacency violations in English, the RCs
and NP-camps, are also subjacency violations in the Romance languages, and
should constitute subjacency violations in all languages. Again, when this
distinction is called for, these subjacency violations will be referred to as
invariant subjacency violations.
Results
Adult learners 1’ersus natives on subjacency
If the accessibility of the subjacency principle is not affected
no significant differences should be found on this principle
by maturation,
between adult
Lnir~erstrlproper-firs
‘jf?
learners of English and native speakers. If it is affected by maturation. adult
learners should perform worse than native speakers on subjacency.
The pattern of performance over the five sentence types for both the native
speakers and the adult learners is shown in Figure 3. This figure combines
the results from NP-camps, relative clauses and WH-camps together, and
presents the number of items correct. Note that the adult learners are performing above chance on the simple and declarative sentences, which means
that they are performing well enough on the rules and structures relevant to
the subjacency principle to be evaluated on that principle.
The adult learners’ performance
on the subjacency violations is significantly below that of the native speakers. Out of 36 items testing subjacency,
native speakers averaged 35 correct and adult learners averaged 21. r(24.4)
= 7.58, p < .OOl. Thus while native speakers find virtually none of the
sentences with subjacency violations to be grammatical,
adult learners find
many of them to be so.
Subjacency
\*ersus no-irl\sersion
It is possible for the subjacency principle to be affected by maturation and
yet have a privileged status with regard to maturation.
To be considered
privileged, the ability to apply the subjacency principle need only be preserved more than the ability to learn language-specific
structures. If the adult
Figure 3.
Mean nwnber correct on 5 sentence types by negative English and Chinese
speakers.
z2
2a-
z
0
26 -
2=
24.
:
2220
18’
-
-
Native
Chinese
I
Simple
OeClsrstiVe
COIWOI
Sentence
No-inversion
Type
Subjacency
learners’ performance on the sentences without subject-auxiliary
inversion (a
language-specific
violation) is worse than their performance
on the subjacency violations. then this possibility. vvould remain viable (though further
analyses would need to be performed to determine whether this is true). If.
on the other hand. the adult learners‘ performance
on the no-inversion sentences is either superior or similar to their performance
on the subjacency
sentences, then subjacency cannot be considered privileged. The latter result
was found. As shown in Figure 3, the adult learners’performed
better on the
no-inversion
sentences than they did on the subjacency sentences, t(22) =
2.21, p < .05, despite the fact that structurally the sentences testing subjacency and no-inversion are very similar. Such a result indicates that the subjacency principle is not privileged in comparison
to this language-specific
structure. Comparisons to other language-specific
structures are made in the
Discussion to determine whether this finding is exceptional.
Subjacency and chance
An important remaining question to answer is whether adult learners have
any hold on the subjacency principle. Do adult learners entirely fail to obey
the subjacency principle, or do they have some tendency to obey the principle, even if not as great as natives ? To answer this, a t test was performed
to determine whether subjects were performing above chance on subjacency.
On subjacency as a whole subjects did perform above chance. r(29) = 2.33.
p < .05 (see the results below for each of the three structure types. NP-camp.
RC and WH-camp, for a clarification of this result).
Subjacency versus controls
Given that subjects are performing above chance on subjacency, it is necessary to rule out the possibility that the result is simply an artifact of a response
bias to say no to sentences of some general type. One way of testing this
possibility is to compare adult learners’ responses to the subjacency sentences
with their responses to the control sentences.
Recall that the control sentences and the subjacency sentences are structurally very much alike. They
both involve wh-movement,
and in very similar structures: hovvever, in the
control sentences the movement is grammatical,
while in the subjacency sentences it is ungrammatical.
Given the similarity of the sentences. one can then
ask whether adult learners are treating them differently,
as they should if
they are observing subjacency rather than responding with a more general
response bias.
As Figure 4 shows, adult learners find the subjacency sentences more unacceptable than the control sentences. The bars in this figure represent the
number of times the subjects respond 110 to the subjacency sentences (for
Figure 4.
Mmn nutnber of no respotrses on subjncetq
Etlglish anti Clkese speakers.
curd control test items b!, turtir~e
36
Chinese
Native
Language Group
which no is the correct answer), versus no to the control sentences (for which
yes is the correct answer). Native subjects correctly reject virtually all of the
sentences with subjacency violations and only occasionally reject a control
sentence. This distinction is much weaker in the adult learners’ performance,
but nevertheless.
the trend is there. Adult learners reject sentences with
subjacency violations more often than they reject the control sentences, t(22)
= 6.08. p < .OOl. A ti’ value of .81 was calculated for the adult learners.
This is a very small discrimination
value, but it is significantly above zero.
t(31) = 6.08, p < .OOl. Not surprisingly, this level of discrimination
is significantly worse than the level of discrimination
found for natives, n’ = 3.89
for natives, t(16.2) = 10.45, p < .OOl, for the difference between them.
The conclusion to be drawn is that there is some remaining tendency in
the adult learner to obey subjacency, and this tendency is not an artifact of
either a response bias or a general belief that wh-movement
cannot occur in
complex sentences. Thus, while subjacency is affected by maturation and is
obeyed less strongly by adult learners than by native speakers, it nevertheless
survrves in a weak and probabilistic form in adult learners.
NP,
RC and WH
The analyses presented so far have used the combined results of the three
structures types, NP-camp, RC and WH-camp.
When these structures are
analyzed individually, all three continue to show evidence of being adversely
affected by maturation. Adult learners’ performance
on the subjacency violations is significantly below that of the natives on each of the three structures:
238
J.S. Jol~r~sor~ and E. L. Nrtvporr
for NP-camp, t( 22.8) = 7.7,~ < .OOl: for RC. ((23.2) = 4.65,~ < .OOl; and
for WH-camp, t(29) = 6.S2, p < .OOl. The fact that none of the structure
types are immune to the effects of maturation is an important one. and one
that will be returned to shortly.
There are differences in the degree to which each of the subjacency structures can be said to be partially available. Above-chance
performance
by
adult learners was only found for the RC structures, r(29) = 6.16, p c .OOl.
Performance
on subjacency for both NP and WH-camp is not significantly
above chance (t(29) = .89, n.s., t(29) = .12, n.s., respectively).
Despite this,
on all three of the structures,
adult learners did distinguish between the
control sentences and the subjacency sentences, though for the NP and WHcamp structures this discrimination is very small: for NP-camp, 8 = .55. t(31)
= 3.05. p < .Ol; for RC, n’ = 1.79, r(31) = 5.2,~ < .OOl; and for WH-camp,
d’ = .9, t(31) = 4.5, p < .OOl.
As this last set of results signifies. subjects were not responding to the
different subjacency test structures in exactly the same way. Adult learners’
performance
on RC subjacency violations is significantly better than their
performance
on both NP-camp and WH-camp subjacency violations, r(22) =
4.96, p < .OOl and t(22) = 5.86, p < .OOl, respectively.
Their performance
on NP-camp and WH-camp is not significantly different. r(22) = 1.34, n.s.
Subjects’ relative ease with RC and their difficulty with NP and WH complements cannot be explained by differentiating
the items in terms of their
status as invariant or parametrically
varying structures.
RC and NP-camp
subjacency violations are both invariant (i.e., both kinds of subjacency violations are ungrammatical
in all languages). while WH-camp subjacency violations are parametrically
acceptable in some languages. All three structures
showed below native levels of performance.
While RC subjacency judgments
were relatively
preserved,
subjacency judgments
in NP-camp were not.
Clearly, subjects are willing to violate even those aspects of subjacency which
are invariant across human languages. This is an important result, because it
means that subjects are violating the outer limit set by the subjacency principle, accepting as grammatical sentences which should be unacceptable
in any
human language.
Non-English parameters for subjacerq
The test was also designed to allow us to determine whether subjects are
obeying subjacency by adopting a different parameter value from English. A
reanalysis using different parameter values will not change the larger conclusion that maturation is affecting subjacency, since none of the English subjacency structure types when tested individually showed such immunity. Since
most of the English subjacency violations are also subjacency violations in
L/nil-ersal properties
239
other languages, any subjects who merely adopted a different parameter
value from that of English would not have performed worse than natives on
all three of the English violations.
For other reasons, however. it may still be of interest to ask whether
subjects show any tendencies of having adopted a different parameter value
than English. Figure 5 shows the pattern of yes responses to each of the six
structures. (The test items here are not scored in terms of number correct.
since correctness ultimately depends upon which parameter value subjects are
adopting.)
The six structures are presented
in the order of least to most
marked, though there are no differences in markedness between the first two
structures or between the last two structures (the order is the same as in Table
2). The ideal patterns for each of the three parameter settings are represented
in Figure 6. It shows the most restrictive setting, the English setting, and the
least restrictive setting, in that order. Although we are only looking for general tendencies,
the actual results clearly bear little resembance
to any of
these idealized results. The obtained pattern of differences and similarities
found between the structures types is unexplained by current theory or understanding of the subjacency principle.
Individual figures were made for each subject’s response pattern over the
six test structures to examine whether individual subjects showed different
patterns which only averaged out to an uninterpretable
pattern. Only a few
Figure 5.
Subjacency
subjacency
parameter setting: mean number
test items by native Chinese.
of yes responses
to control and
IL
11 10 96.
765.
43.
21 0
H
•A
0
p
NPControl
WHSubjacency
NPSubiwency
RCSubjeoency
2-m
J.S. Johnsor
Figure 6.
and E. L. ,\‘eic,port
Ideal pattern of yes responses for the most restrictiir,eparameter
setting
(S. S’). for the English parameter setting (S. NP), and for the least restrictive
setting (S’. NP).
Most
Restrlctlve
English
Parameter
Least
Restrictive
Setting
subjects could be identified as having adopted any particular parameter value,
and only on a probabilistic basis. All of the remaining subjects could not be
identified as having adopted any of the parameter values. Thus, although the
average pattern of the group is not typical of all subjects, it nevertheless is a
composite of the basic facts that have been found for the adult learners.
Validating the subjacency
measttre
While running Study 1 (and 2), as it became clear that subjects were violating the subjacency principle, a natural question arose: are subjects really
violating subjacency when they accept the subjacency violations as grammatical? Or is it the case that, by some process different from that of native
speakers, subjects are interpreting
the wh-question item as moving from a
position which does not violate the subjacency principle? If the latter is true,
then it would not be subjacency that they were violating, but rather a number
of language-specific
properties simultaneously.
A brief test was devised to
evaluate this possibility and was given to a subset of the subjects following
the grammaticality
judgment test.
The test was designed to ask whether subjects were correctly interpreting
questions with subjacency violations. Our analysis of the crucial test sentences
as violating subjacency assumes that the wh-word is extracted from the embedded clause of the sentence. If subjects parse these questions in that way,
then they should be able to answer those questions given appropriate information. On the other hand, if they are parsing these questions in a different
way. then their attempts to answer them should reflect this error. The test
was a paper and pencil test which contained 1s test items. each made up of
a declarative sentence and a related question. Subjects were asked to read
each declarative sentence and its related question. and answer the question
based on the content of the declarative sentence. They were told explicitly
that some of the sentences would be ungrammatical.
but that they should
nevertheless attempt to answer the question. Subjects were asked to respond
in one of three ways: (1) if they believed that they knew the correct answer.
they were to simply write that answer down: (2) if they had no idea what the
question was asking. they were to write “I don’t know”: (3) if they knew what
the question was asking, but didn’t believe that the answer was given in the
declarative sentence. then they were asked to write “something/someone
not
mentioned”.
This latter option was essential for this test, for it allowed subjects to indicate if they believed that the wh-item was extracted from somewhere other than the intended position. For esample. if the subjects were
interpreting
the sentences with subjacency violations as wh-movement
in the
upstairs clause, then this last response possibility allows them to report that
belief.
Twelve of the test questions involved subjacency violations (4 of each of
the three structure types NP. RC and WH): the other 6 questions were
control sentences used as fillers in this test. An example of one of the test
items is shown in (3) below:
(3)
The policeman knew where the thief hid the jewels.
What did the policeman know where the thief hid’?
The correct answer to this question would be jeivefs.
One of the filler control sentences was constructed so that it did not provide
an answer in its declarative match. This sentence was included to encourage
subjects to use the response option “someone/something
not mentioned”
whenever
they felt that the WH-item was not moving from out of the
downstairs clause, and also to serve as a check to see whether they felt
comfortable
using that option. This test item is presented in (4) below:
(4)
Ted asked where he could find Carol
Who did Ted ask where he could find Carol?
A similar sentence-question
pair was given as a sample item before the test,
along with two other normal grammatical test items. To make sure the subjects understood all that was expected of them, all of the sample items were
worked through orally with the subject.
If subjects accepted subjacency violations because they mistakenly thought
the wh-item was moved from a position which did not violate the subjacency
principle, then their responses should primarily be of the form “something/
212
J.S. Johnson
and E. L. A’rwport
someone not mentioned”.
If, on the other hand. subjects violated subjacency
while fully knowing where the wh-item was moved from. then their responses
should primarily be the correct answer to the question. A response of “I don’t
know” supports neither contention.
When subjects respond “I don’t know”.
they neither believe that the wh-item came out of the matrix sentence nor do
they believe that it came out of the embedded sentence.
Seven Chinese adult learners and six native speakers took this test. The
results from the adult learners are presented in Table 3(a). which shows the
average number of responses made out of a total of 4 possible responses for
each structure type (the rows do not always add up to 4 due to rounding).
Though not shown in Table 3, it should be noted that all of the subjects, both
adult learners and native speakers, used the response option “some person
not mentioned” on the control sentence designed to provoke that response.
This suggests that subjects would have used this response when they thought
it was appropriate.
Column 2 in Table 3(a) nevertheless shows very little use
of this response to the target test items. There is. therefore,
little indication
that subjects are incorrectly parsing the questions with subjacency violations.
A comparison of the results in columns 1 and 3 shows that subjects in fact
generally know the correct answer for the subjacency violations: this effect
is strong for NP-camp and WH-camp structures. In contrast, for quite a few
Table 3(a). Mean responses of Chinese adult learners on subjacency
comprehension
test
~_.~
Structure type
Correct
Unmentioned
Doesn’t know
0.28
NP
3.57
0.1-I
RC
2.57
0. I4
I .28
WH
3.28
0.28
0.43
Note: maximum possible for each structure type = 1.
“One wrong answer was included in this category because the answer. though wrong. still
violated subjacency.
Table 3(b). Mean responses of native speakers on subjacency
Structure type
Correct
comprehension
test
Unmentioned
Doesn’t know
NP
3.83
0
0.16
RC
1.16
0
2.83
WH
3.00
0
1.00
Note: maximum possible for each structure type = 4.
L’rriversnf properties
243
of the RC structures subjects are unable to interpret the question. It may be
that their inability to interpret these violations is an indication of the accessibility of subjacency in this case. This would correspond to the fact that subjects are best at rejecting RC subjacency violations.
The results from the native speakers are presented in Table 3(b). As shown
in column 2, the native subjects never interpret the wh-item to be moving
out of a position other than the correct one. Comparing columns 1 and 3,
however, it is evident that subjects do not always know what the correct
answer is. Like the non-native speakers, they generally know the correct
answer in the NP-camp and WH-camp structures. In the RC structures, however, just the reverse is true. They rarely know the correct answer in these
cases. Non-native subjects in fact appear to be better than natives at understanding questions which contain RC subjacency violations. Again, this inferiority in interpreting the RC subjacency violations could be due to the fact
that native speakers strongly observe subjacency,
and consequently
find it
even more difficult to figure out what the sentences mean.
The results of this exercise suggest that when subjects accept NP-camp and
WH-camp subjacency violations, they are indeed violating the subjacency
principle; they do not appear to mistakenly believe that the extraction came
from a different and grammatically
permissible site. When subjects accept
RC subjacency violations, it is less clear that they are definitely violating the
subjacency principle: however, this issue is less important for RC subjacency
violations, since on this structure subjects do not often violate subjacency.
Given these results, the results and conclusions presented above for the grammaticality judgment task will be assumed valid.
Discussion
The various analyses of this study converge to suggest that the subjacency
principle is not fully accessible to the mature learner for the learning of a
second language. The basic finding of non-native performance
on subjacency
by adult learners accords with the results of other studies which found similar
results using a written evaluation of subjacency (Bley-Vroman
et al., 1988;
Schachter, 1989). The agreement between these studies and the present one
not only supports the earlier studies’ findings, but also suggests that the
results obtained were not due to extragrammatical
properties particular to
any one set of stimuli.
There also was no evidence to suggest that subjacency is spared relative to
language-specific
structures.
Subjects’ performance
on questions with subject-auxiliary
inversion violations, a language-specific
property,
was compared to their performance
on questions with subjacency violations. Rather
than performing
better on subjacency than on subject-auxiliary
inversion
violations, they performed
worse. This result is also supported by a more
general comparison
with the data from J & N. In the present study adult
learners’ performance
was not above chance on two of the three subjacency
structures (NP-camp and WH-camp). In contrast, in the 12 language-specific
structures tested in J & N, only two were not above chance in adult learners.
Adult learners’ error rates on the 10 above-chance structures in J & N ranged
from 10% to 40%. In comparison,
the error rates on the three subjacency
structures in the present study are 22%. 45% and 51% (for RC, NP-camp
and WH-camp subjacency structures respectively).
Thus, except for the RC
structures, the subjacency results are like those language-specific
structures
which were the most difficult for Chinese adult learners in J & N. Undoubtedly there are ways of creating language-specific
test items more difficult than
those tested in J & N. The point made here is not that subjacency is more
difficult than language-specific
structures, but rather that subjacency is not in
general less difficult than language-specific
structures. In short. subjacency is
not spared from the decline found for language-specific
structures.
Although maturation
has affected the status of subjacency.
the results
suggest that there may nevertheless be some remaining tendency to obey the
principle, if only probabilistically.
This was evident with relative clauses, but
much less so with NP and WH-camp structures since performance
was not
significantly above chance on those structures. Despite this, subjects did distinguish in their responses between subjacency violations and control sentences on all three structures, suggesting some minimal discrimination
between control sentences and subjacency violations in all instances. The probabilistic observance of this constraint corresponds
to the probabilistic observance of language-specific
rules by mature learners in previous studies, both
for second language learners (Johnson & Newport, 1989) and first language
learners (Newport & Supalla, 1990). The data now show a fragile knowledge
of both grammatical rules and constraints. It may seem odd to suggest that
adult learners have partial access to subjacency (just as it seems odd to suggest
that they have partial access to a language-specific
rule). In both cases, however, this odd, probabilistic status is what the empirical data suggests. How
this partial knowledge should ultimately be characterized
is unknown at this
time.
Contrary to what some researchers have suggested (duplessis et al., 1987:
White, 1984a), subjects’ difficulty in obeying subjacency was not the result of
merely having set the principle’s parameters incorrectly.
Incongruent
with a
parameter-setting
approach, none of the response patterns predicted by the
three possible parameter values found in languages of the world were evident
in the results. More importantly,
the pattern of subjects’ performance
was
Univrrsd proprrties
743
not accounted
for by distinguishing
between
universally
invariant
and
parametrically
varying aspects of subjacency.
If subjects resisted violating
subjacency structures which are invariant across languages, they would have
performed
well on subjacency violations in both relative clauses and NPcamps. While some reluctance was shown to accept RC subjacency violations,
little such reluctance was shown to accept NP-camp subjacency violations.
Given this. one cannot rule out the extraordinary
conclusion that adult learners of a language will sometimes form hypotheses or rules in that language
which are unnatural to human languages.h
In conclusion, Study 1 showed that the subjacency principle is deeply affected by maturation.
This maturational
effect was shown not merely to be
the result of an inability to set the parameters of this principle. Indeed, the
entire integrity of the principle appears to be affected by maturation.
Upon
maturity, then, language learners show only weak tendencies to observe linguistic universals, and may therefore implicitly form hypotheses about their
late-learned
language that violate these universals.
Study 2
Study 2 was designed to extend our understanding of the critical period effects
on subjacency by observing how subjacency manifests itself over varying ages
of acquisition of the language. Given that effects of maturation on subjacency
have been found for adult learners, it is of interest to know what the pattern
of performance looks like over a range of ages of acquisition. Questions which
need to be answered include the following:
(1) Does the ability to obey the subjacency principle undergo a gradual
deterioration
over age of acquisition similar to what has been found for language-specific details? Or is the ability to obey subjacency completely spared
from maturational
effects until some particular maturational
point (e.g., puberty), dropping abruptly to the adult learners’ level of performance?
(2) If maturational
effects do appear before adulthood,
at what age do
such effects begin to appear? Is the timing of the decline the same as for
language-specific
structures or is it different?
(3) If maturational
effects of some kind are found, is there a qualitatively
?-his is obviously a strong claim to make. It crucially depends on the correctness of the cross-linguistic
facts stated for subjacency. For example. if it is later found that some languages do violate this outside setting.
however exceptional they may be. then clearly violating that setting is within the set of natural hypotheses
about language (see footnote 4).
different pattern of performance
than that found for adult learners (indicating
that perhaps subjacency is differently affected before and after maturity), or
are younger learners performing
in a qualitatively
similar manner as adult
learners?
Methoii
Subjects
Again, native speakers of Chinese were asked to participate to provide
continuity between the investigation of the adult learners from Study 1 and
the investigation of age effects here. This allowed for comparisons to be made
between the results of J & N and the present study. and. as in Study 1. it
prevented
the possibility that subjects perform well on subjacency due to
explicit or superficial transfer from their own native language.
Subjects, then, were 21 native Chinese speakers who learned English as a
second language before adulthood. The primary basis for selecting subjects
was the age at which they moved to the U.S.A. and thereby first became
immersed in English. Subjects were selected to form a representative
sample
of young learners varying in age of arrival in the U.S.A. from 4 to 16 years
of age.
The resulting 21 subjects consisted of 9 males and 12 females. The distribution of the sexes was fairly even across age of arrival. At the time of test,
subjects were between the ages of 17 and 25, with a mean age of 20. Years
in the U.S.A. ranged from 5 to 1.5, with a mean of 9.6 years. There was some
tendency for younger arrivals to have more years of experience in the U.S.A.
than the older arrivals: however, in previous studies with the same-age learners (Johnson & Newport, 1989; Oyama, 1978) no effects of length of experience have been found after 5 years of immersion.
Subjects were recruited from the university population
to ensure ample
exposure to English, and to ensure comparable
social backgrounds
among
themselves, as well as between them and the adult learners described in Study
1. Although learners from Study 1 and the present study were all selected
from a university population,
there nevertheless
were differences between
them. While the adult learners were selected from the population of graduate
students, post-doctoral
fellows and faculty, the younger learners were primarily selected from the undergraduate
population.
with the exception of two
graduate students. There are thus some possible differences between the two
groups in terms of the type of university experience
they received and, of
course, in terms of the amount of education they had at the time of test, with
the adult learners being more highly educated than the younger learners.
Subjects were recruited through letters, sign-ups and by word of mouth.
L’tlirwrcrl
proprrtirs
237
The language environment
of the young arrivals was very similar. Before
they came to the university, they spoke Chinese in the home to their parents
and English at school. This is similar to the adult learners’ environments
in
terms of the amount of contact with both Chinese and English. After they
came to the university, most of them spoke primarily English. having only a
few friends with whom they could speak Chinese. When they went home to
their Chinese families living in the U.S.A. for the summer and other vacations, they returned to speaking Chinese in the home, and English outside
of the home.
In contrast to the adult learners of Study 1, none of the subjects of this
study had English classes before arriving in the U.S.A.
No subjects were eliminated, either due to their performance on the simple
questions, or for any other reason.
Procedure
The procedure
used in this study was identical
to that used in Study 1.
Materials
The materials
used in this study were identical
to those used in Study 1.
Results
To allow for comparisons,
the results of the natives and adult learners from
Study 1 will be included in many of the analyses. Also to facilitate comparisons, for most of the analyses the younger learners will be grouped by age
of arrival into three categories: 4-7 (n = 6), 8-13 (n = 9), and ltl6
(n =
6)-year-old
arrivals. The cut-offs for these groupings of age of arrival are
similar to those used in past research (Johnson & Newport, 1989: Snow &
Hoefnagel-Hohle,
1978).
Subjacency
over age of arrival
The primary purpose of this study was to determine the pattern of performance on subjacency over age of arrival. Does the ability to obey subjacency
continuously
decline over age of arrival, as had been found by J & N for
language-specific
rules, or are there important discontinuities
in its accessibility over age?
Figure 7 presents performance
on subjacency (see the bold line) for all of
the different age groups, including the natives and adult learners from Study
1. As can be seen. there is a fairly continuous decline in the observance of
subjacency as age of exposure to the language increases. This continuous
decline is supported by a significant linear trend over age of arrival (F( 1,50)
24s
J.S. Jol~r~son anti E. L. h’e~~port
Figure 7.
The relariotrship berlveetl age of arrir3al in the U.S.A.
on wbjacetyv
test items and no-inr,ersiorr test items.
anti tlrttnber correct
““I
I
3432 30 28 26 24 22 .
20 .
---*--
SubJacency
Nblnverslon
I
18
Nalive
4 10 7
8 10 13
Age
14 10 16
Adult
of Arrival
= 25.95, p < .OOl. on a linear trend test with unequal intervals). and a
corresponding
significant correlation
between age of arrival and subjacency
score (Y = - .63, p < .OOl). While the 4-7-year-old arrivals are not significantly
different from the natives on subjacency (r(6.2) = 1.49, p < .lO with a onetailed test),’ the remaining groups are (native vs. 8-13, r(9.9) = 2.45, p <
.05; native vs. 14-16. t(5.3) = 3.05, p < .05), all using a two-tailed test).
In sum, there are no sharp discontinuities
in subjects’ ability to obey the
subjacency principle as age of arrival increases. Rather, this ablhty undergoes
a gradual decline with age of arrival. similar to what was found for languagespecific rules in J & N.
To rule out the possibility of a qualitative change in performance occurring
before adulthood, it is still important to ask how much of subjacency remains
in the young learner and what the subjects’ pattern of responses is across all
sentence forms given in this test. In presenting these results the organization
of the presentation
of Study 1 will be closely followed.
Subjacency
I’ersus no-inversion
In Study 1 we asked whether the subjacency principle. though affected by
maturation,
still has a privileged status with respect to it. Here we consider
‘The apparentdecline in the J-7 group is beins carried hy one 7-year-old subject: uithout that subject the
menn is J-1.
L’rri~~cr.strl
propcrrirs
this question again by comparing performance
over age of arrival on subjacency to that on no-auxiliary inversion sentences. Figure 7 shows subjects’
performance
on no-inversion sentences over age of arrival as a dashed line.
At none of the age groups is the difference between subjacency and no-inversion sentences significant (for the -1-7 group, r(7) = .17. n.s.: for the S-13
group, r(8) = .7, n.s.: for the 1-I-16 group. r(5) = .06. n.s.). Compared to
language-specific
rules, then, the subjacency principle is not privileged with
respect to maturation
at any age of acquisition. Subjacency and languagespecific rules show comparable declines with age of acquisition.
These results appear interesting for another reason. In Study 1, the adult
learners were shown to perform worse on subjacency than on the no-inversion
sentences. With the younger learners this difference does not exist. It appears
that subjacency undergoes a greater decline after age 13 than does no-inversion. From age 8-13 to adulthood, subjacency drops 27.53 percentage points
on the test. while no-inversion drops only 10.92 percentage points. The potential significance of this discrepancy will be discussed in the Discussion.
Subjacency
and chance
All of the younger learning groups are performing above chance on subjacency (for the 4-7 group. t(S) = 12.99. p < .OOl; for the 8-13 group. r(8) =
11.72, p < .OOl; and for the 14-16 group, r(5) = 3.86, p < .05). Unlike what
was found for the adult learners, the individual results for each of the three
structure types, NP-camp, RC and WH-camp, all show above-chance
performance as well.
Subjacency
versus controls
To rule out the possibility that subjects are performing above chance on
subjacency due to a response bias to say “no”, subjects’ “no” responses to
control sentences (which would be an incorrect response) were compared to
their “no” responses to the subjacency sentences (which would be a correct
response). Figure 8 shows this response difference. For comparison, the adult
and native responses from Study 1 are included in this figure. In each of the
groups, subjects distinguish between control sentences and subjacency sentences; for the 4-7 group, t(5) = 12.83, p < .OOl; for the 8-13 group. r(8) =
16.37, p < .OOl; and for the 14-16 group. t(5) = 4.35, p < .Ol. However, as
can be seen in the figure, the degree of discrimination
between control and
subjacency sentences continuously declines as age of arrival increases.
The discrimination
between control and subjacency sentences is most directly quantified by the d’ values for each of the age of arrival groups. Including the native and adult learners from study 1, these d’s are as follows: for
natives, d’ = 3.89, for 4-7. d’ = 2.7; for 8-13, d’ = 2.6: for 14-16, d’ = 1.35;
‘50
J.S. Jolrmotl
Figure S.
trtiti
Mean tlutnber
var_ving
E. L.
of
Xe~c~~~o,.r
no
respotlses
otz srrbjmenvy
atld
control
test items
b>
rages of orrird
36
n
Subjacency
II
Control
6
Native
4 to
a
7
Phge
t0
of
13
14
to
16
Adult
Arrival
and for adult learners, n’ = .S-t. Like the overall subjacency scores, the ability
to discriminate subjacency from control sentences declines continuously over
age of exposure (F(lS0)
= 91.72. p < .OOl. on a linear trend test with
unequal intervals).
Each of these ti’ values is above chance (for native and adult learners, see
Study 1); for 4-7. r(5) = 11.69. p < .OOl: for S-13, r(S) = 17.2. p < .OOl; for
14-16. r(5) = 3.59, p < .05. However. only the younger age groups show
high levels of discrimination.
The n’ measures are useful for another reason. While the overall subjacency score showed no significant difference between natives and the 4-7year-old arrivals. the more sensitive 8 measure does show a significant difference between these groups (r(1l.l) = 3.46. p < .Ol). This suggests a decline
in expertise with age of exposure beginning as early as age 4-7.
NP, RC and WH
For the adult learners, RC subjacency structures were easier than the WHcamp and NP-camp subjacency structures. No differences among the three
structures appear in any of the three younger age groups due to the small A
in each of these groups. However. the same advantage for RC subjacency
structures over the other structures appears when these three age of arrival
groups are combined. Out of 12 possible. the means for the combined age of
arrival group are as follows: NP = 9.67: RC = 11.14: and WH = 10.04. Both
C’rtir~erstrl properties
Figure 9.
Subjacency
parameter
subjacency
test items
setting:
mean
ntrrnber
of yes respomes
to control
'i[
__
and
by age of arrival.
12
11
s
10
E
9
i
L
B
0
>
6
7
6
6
L
4
$
3
:
2
1
0
4
to
7
6
Age
to
13
14
to
16
of Arrival
NP and WH differ significantly from RC (t(20) = 2.S2, p < .05, ((20) = 3.06,
.Ol, for NP vs. RC and WH vs. RC. respectively).
whereas NP and WH
do not differ significantly from each other (t(20) = .63, n.s.). As in the data
for adult learners, then. the younger learners’ pattern of performance
on
subjacency does not suggest a contrast between invariant (RC and NP-camp)
and parametrically
varying (WH-camp) aspects of subjacency.
p <
Non-English
parameters
for subjacency
Recall that the pattern of “yes” responses to the control sentences and the
subjacency sentences should inform us as to what parameter value subjects
are tending to obey. Subjects’ yes responses to the control and subjacency
sentences are presented
in Figure 9, for each of the three age of arrival
groups. If subjects adopted the English parameter value, the first three structures should have received a yes response to a high degree, and the last three
structures to a low degree. This is obviously the case for the 4-7 group and
the 8-13 group. On the other hand, the 14-16 group, like the adult learners
of Study 1, show a pattern not well described by any known parameter values.
For this group, NP-camp control sentences (the third bar in the figure) show
a chance response level, making their overall response pattern unlike either
an English or a non-English parameter setting.
Validating the subjacency
In Study
measwe
1. a subset of the subjects
were given an additional
test to see
32
J.S. .Joi~nson rlnti E. L. .l’elr~po,_t
how they interpreted questions with subjacency violations. Our main concern
was to make sure that subjects were not interpreting
the wh-item as moved
from a position which did not violate the subjacency principle. Subjects read
declarative sentence and question pairs. and were asked to answer the question based on information from the paired declarative sentence (see Study 1
for details).
A sample of the younger learners took this test as well. For brevity. their
responses have been collapsed across structure type and are presented in
Table 1. (Similar relative differences
between the structures found in the
adult learning group occurred here as well.) The N for each group is very
small: 6 for the -4-7 group, 2 for the 8-13 group and 2 for the 1-l-16 group.
In these tables. “correct” responses mean that the subjects were able to
correctly answer the question containing the subjacency violation, indicating
that they interpreted
the questions as intended. Responses such as “something/someone
unmentioned”
indicate that the subject is extracting the whitem from some place other than what was intended, possibly from a place
that would not violate subjacency.
If subjects primarily made responses of
this sort, then their errors on the subjacency violations in the grammaticality
judgment task should not be interpreted
as subjacency violations. “I don’t
know” responses cannot be interpreted
to support either position.
Table -l shows the mean number of each response out of a total possible
of 12. Subjects in all three age of learning groups were generally able to
correctly answer the questions with subjacency violations. They rarely responded that they could not understand the question, and only occasionally
did they believe the wh-item was being moved from a place other than what
was intended. For this group of arrivals. then, we can take the results of the
grammaticality
judgment test on all of the subjacency structures as valid.
All of the subjects used the response option “some person/thing not mentioned” on the appropriate
control sentence designed to provoke that response. This indicates that subjects’ lack of using this response on the subjacency questions was not due to a bias to avoid that type of response.
Table 4.
Mean responses
test
of the different aged arrivals on subjncenc!
Age of arrival
Correct
1-7
Y .(X1
s-13
Y.50
I-l-16
11.00
Note:
maximum
possible
overall
=
I?.
cotnprel~ension
Uni~rrstrl propertirs
It may be interesting to note that the -l-7-year-old arrivals. like the natives.
are the least able to answer questions with subjacency violations. Indeed. as
age of arrival increases, subjects appear generally better at answering the
questions with the subjacency violations at the same time that their performance on grammaticality
judgments on subjacency sentences decreases.
Most importantly,
these results show that when subjects err on subjacency.
they are not doing so because they incorrectly interpreted
the wh-item as
moving from a position that would not violate subjacency. This result thus
reaffirms the validity of the conclusions formed in study 2, by verifying that
subjects are parsing the sentences with subjacency violations correctly.
Discussion
The results show a gradual decline in the accessibility of subjacency similar
to what has previously
been reported
for language-specific
structures
(Johnson & Newport, 1989; Oyama, 1978: Patkowski, 1980). This was evident in subjects’ ability to reject subjacency violations, as well as in their
ability to discriminate
between subjacency
violations and the similar but
grammatical
control sentences.
By either measure, performance
declined
gradually with increasing age of arrival.
Moreover, as in Study 1 there was no evidence to suggest that subjacency
was in any way spared from maturational
effects relative to language-specific
structures. Before the arrival age of 13, performance
on subjacency and on
no-inversion
showed precisely the same declines over age of arrival. After
that age, performance
on subjacency underwent a greater decline than no-inversion. From this last result it may be tempting to conclude that universals
are different from language-specific
properties in that the former is even more
vulnerable to maturation.
However, it is more likely that it reflects a difference in test item difficulty. The same type of abrupt decline was found for
the most difficult structures tested by J & N (e.g., determiners).
Thus in
general, test items which are more difficult undergo faster rates of decline
over maturation.
The different rates of decline in the present results likely
reflect a general difference in test item difficulty rather than a fundamental
distinction between universal and language-specific
properties.
The d’ values suggest that subjacency becomes less accessible to the learner
as early as between the ages of 4-7; previous studies on second language
acquisition have only found declines after this age. Again, we believe this
reflects the level of difficulty of the test and the sensitivity of the d’ measure
(only the d’ measure picked up this earlier decline). A similar decline at 4-7
has been found in primary language acquisition for the more difficult structures of the target language. Newport and Supalla (1990) have found subtle
subtle but consistent declines in the morphology of American Sign Language
with deaf individuals whose first exposure to the language occurred between
the ages of -1 and 7. These results suggest that. with sensitive measures. small
declines may be detected as early as -1-7. The onset of this maturational
change. then. may occur at an earlier age than previous data had suggested.
One of the more important findings of Study 1 was that adult learners
sometimes
failed to reject the invariant subjacency
violations which are
thought to be outside of the range of possible human languages. Performance
on these structures,
like on other structures.
declined over age of arrival.
Evidently,
as age of learning increases learners of a second language will
sometimes violate the universal boundary set by the subjacency principle and,
by implication. internalize a grammar which allows structures thought to be
unnatural to human languages.
General discussion
Together,
these studies suggest that linguistic universals such as subjacency
become less accessible to the language learner with increasing maturation. In
this sense, subjacency does not differ from language-specific
structures: both
are affected by maturation. apparently even to the same degree and manner.
The most general import of these findings is to make even stronger the
claims of prior literature showing that there is a critical or sensitive period
for the acquisition of language. First. there are the results from first language
acquisition (English: Curtiss, 1977. 19SS: American Sign Language: Newport
& Supalla, 1990). Second there are the results on second language acquisition
of language specific structures, each involving different measures of grammatical proficiency (Oyama, 1978; Patkovvski. 1980; Johnson & Newport, 1989).
And now there are the results from several studies, including this one, which
demonstrate
critical period effects for language universals as well (Schachter,
1989; Bley-Vroman
et al., 1988).
As noted at the beginning of this paper. we have not distinguished between
the terms “critical” and “sensitive” period, instead using the term “critical
period” for the general phenomenon
of declining competence
over maturation. As many investigators
have noted,
virtually
all such observed
phenomena
in the animal literature (e.g., imprinting) are characterized
by
graded rather than abrupt behavioral changes over age, with the boundaries
determined primarily but not exclusively by maturational
factors (Immelman
& Suomi, 1981). Our data, and the prior hterature on both first and second
language acquisition, are of this same graded character. and thus in this sense
show what some may prefer to call a sernitir.e period rather than a critical
period. Some declines in the ability to learn language are visible as early as
age 4-7. with increasingly pronounced
effects up to adulthood.
Moreover.
like some but not all such phenomena in the animal literature, for language
acquisition there appears to be some residual. though greatly diminished,
abilitv to learn even during adulthood. Whether this residual ability is due to
the diminished operation of the same faculty, or rather to the operation of
other learning procedures,
is unknown (cf. the literature on imprinting for
controversy on the same issue).
As in other areas in which critical or sensitive periods have been observed.
the demonstration
of such a maturational effect does not uniquely specify the
nature of the underlying mechanism (Bateson. 1979; Hess, 1973; Newport,
1991). One possibility is that there is a domain-specific
language faculty which
is itself the locus of this maturational
change. An alternative possibility is
that, while there may (or may not) be a domain-specific
language faculty, it
is not the locus of this maturational change: rather, maturational changes may
occur in other mechanisms which impact the way. languages are learned late
in life. In any case, the demonstration
of a critical or sensitive period for
language acquisition does provide important evidence for the more general
claim that native language learning operates successfully by virtue of a set of
internal constraints, at least some of which are present only early in life.
What are the further implications of the finding that language universals
and language-specific
details do not differ with regard to age of learning?
Here we must separate two types of potential implications. First, we do not
believe that such a finding bears closely on the question of whether theories
of nnrive languages or their acquisition should entail such a distinction. Linguistic arguments abound for the relevance of such a distinction in the formal
description of natively acquired languages (though theories of course differ
in how this distinction is embodied).
Our own data bear only on whether
these aspects of language change together or separately over maturation. We
have argued that they change together, thus entailing some important implications for theories of non-native (i.e., late-learned)
languages and their acquisition.
As noted at the outset of this paper, several recent theories of second
language acquisition have claimed that adult learners have full access to universal grammar, and only have difficulty in properly setting or resetting the
parameters
of UG (Flynn, 1983, 1984; Liceras, 1985; Mazurkewich,
1984;
Phinney, 1987; Ritchie, 1978; White, 1984a, 1984b). One virtue of such a
theory is that it conceptualizes
the adult learning process as precisely like that
of child language acquisition, differing only in one (theoretically
coherent and
also theoretically less significant) component.
However, our data, along with
those of Bley-Vroman
et al. (1988) and Schachter (1989), are at odds with
such theories; the changes that occur between childhood and adulthood in
language learning seem to affect all aspects of grammar acquisition, including
access to UG. the ability to set parameters.
and the ability to master the
language-specific
details of the grammar. One possibility is simply to weaken
the claims above: perhaps adult learners have some partial access to all of the
mechanisms of native acquisition.
but all operate in a weakened or more
errorful fashion. If the changes that occur with maturity are in this sense
quantitative
and not qualitative, an enterprise which studies the probabilistic
observance of universals, markedness and transfer effects would still be viable. A quite different possibility is that. over age. the mechanisms of native
acquisition are gradually replaced by entirely different, and less constrained,
learning mechanisms.
In any case. theories which hypothesize that much of
adult language acquisition comes whole and intact from native language acquisition will apparently
require modification,
and perhaps theories which
hypothesize
much more broadly different learning mechanisms for adults
must be developed.
It has been said that children bring to the first language task two types of
information:
one, an innate set of constraints on what human languages may
be like, which rule out certain kinds of structures or grammars as impossible
in any human language: and, two. a set of procedures
for using linguistic
input to determine the particular features of the target language among the
range of possible forms human languages may take. In contrast to this view
of the child learner, the present and prior results suggest that the older learner
has both a weakened or diminished set of universal constraints on human
languages in general, and a weakened or flawed set of procedures for delineating the particular features of the target language. In the older learner
these sources of knowledge appear to operate only probabilistically,
and overall at much lower levels of accuracy. The outcome of late second language
learning, then, is an acquired language which, though still probabilistically
similar to the target language. is imperfectly mastered and sometimes even
violates universal constraints on human languages.
The breadth of critical period effects found first for language-specific
structures, and now for structures considered to be governed by universal principles, suggests that whatever the nature of the endownment
that allows humans to learn languages, it undergoes a very broad deterioration
as learners
become increasingly more mature. An account of this decay would therefore
have to entail a general maturational
change, through a mechanism broadly
affecting all aspects of language acquisition. It remains for future research to
say what exactly the initial endownment
is and why it declines with age.
L’rlivrrsnl
proprrtirs
‘57
References
Allwood. J.S. (1976). The complex np constraint as a non-universal rule and wme semantic factors influencing
the acceptability of Swedish sentence, which violate the CNPC.
Prtpers itt Littgttistics.
( 1979).
Bateson. P.
Bley-Vroman.
Vol.
(irtiwrsit~~ of:Mossoch~rsrrr~ Occ~rriotrd
I
How do sensitive periods arise and what are they for’? Attimtl
R.. &z Chaudron.
C. (lY87).
Behmior.
Review essay. A crlrique of Flynn’s parameter
27. 471U-186.
setting model of
second language acquisition. Submitted for publication.
Bley-Vroman.
R.. Felix. S.. Sr loup. G. (IY8Y). The accessibility of universal grammar in adult language
learning. &co4
Lojlgwyr
Research. 4. 1-X.
Chomsky. N. (lY68).
Ctrtxrtt isstresirt lirtgtristic theor?. The Hague: Mouton.
.-lsprctsoft/w t/tro,:\ ofsuttrr.~.Cambridge. MA: MIT Press.
Lortywgr ottd mid.
New York: Harcourr Brace Jovanovich.
Chomsky. N. ( IYSI).
Lecttrrer ott yo~wwt~etrt
Chomsky. N. (IY6-4).
Chomsky. N. (IY65).
ortd bittdirty.
Dordrecht:
Foris Publications.
Clahsen. H.. Sr Muysken. P. (lY86). The availability of universal grammar to adult and child learners: A study
of the acquisition of German word order. Second Lotlgwye
( 1977).
Curtiss. S.
Gntie: A p~~cholittytristicrttr~~
Curtiss. S. (1988).
Research. 2.
ofcr modent do! “wildchi/d”.
New York: Academic Press.
The case of Chelsea: A new test case of the critical period for language acquisition.
Unpublished manuscript. UCLA.
duplessis. J.. Solin. D.. Travis.
L.. & White.
L. (lY87).
UG or not UC.
that is the que>tion: A reply to
Clahsen and Muysken. Srcorrr! Ltrnyltnge Resrctrch. 3.1.
Erteschik.
Flynn. S.
N. (lY73).
On the nature of island constraints. MlT
(1983).Differences
sal grammar.
Dissertation.
June lY73.
between first and second lanpuape acquisition: Setting the parameters of univer-
In D. Rosers & J. Sloboda (Eds.).
.-\cqttisitiotr
of sytbolic
rkills.
Zcu
York:
Plenum
L.H.
Bell. &
G.W.
Barlow.
Press.
Flynn. S. (lYY-4). A universal in L2 acquisition based on a PBD
Cttiwrrsct/s
itt secorttl
D. Nelson (Eds.).
Gold. E.M.
(1967).
Langua_ee identification
Hess. E.H.
(1973).
Impritttitty.
Immelmann.
Rowley: Newbury.
in the limit. Iufonnotiott
ottd Corrtrol
10. 417474.
New York: Van Nostrand.
K.. & Suomi. S. (1951).
Sensitive phases in development.
L. Petrinovich. & M. Main (Eds.). Behoviorol
Johnson. J.S.. & sewport.
typology. In F. Eckmsn.
ktrtytrnge ncqttiritiorr.
E.L.
development.
In K. Immelmann.
Cambridge: Cambridge Cniversity Press.
(1989). Critical period effects in second language learning: The influence of
maturational state on the acquisition of English as a second language. Cogttiti\v
Krashen. S.. Scarcella. R.. & Long. M. (Eds.) (1982).
Child-o&r/t
differettces
Psyhologx.
21.60-YY.
bt second lottgttoge ocqttiritiott.
Rowley. MA: Newbury House.
Lenneber_e. E. (1967).
Liceras. J.M.
(1955).
periphery.
Mazurkewich.
Biologicnlforrtt~~tiotts
L2 learnability:
Unpublished
I. (IYY4).
oflottgttoge.
Delimiting
manuscript.
Dative questions and markedness.
Ltttiversols itt second lnrtgrtoge ocqtrisitiorr.
Newport.
E.L.
(Eds.).
Newport.
E.L..
New- York: Wiley.
the domain of core grammar as distinct from the marked
In F. Eckman.
L.H.
Bell. 8; D. Nelson (Eds.).
Rowley: Newbury.
(1991). Contrasting conceptions of the critical period for language. In S. Carey b: R. Gelman
Epigertisis of mittdc Essays on biology
nrtd cognitiott.
Hillsdale. NJ: Erlbaum.
& Supalla. T. (1990). A possible critical period effect in the acquisition of a primary language.
University of Rochester. manuscript.
Otsu. Y. (1981).
.5tti~ersol
grommor
ottd syrmctic
developmmt
itt children:
Toword
o theory
of symctic
de~~elopmetrt. Doctoral dissertation. MIT.
Oyama.
S. (1978). The sensitive period and comprehension of speech. Workittg
I-17.
Papers on Bilittguolirm.
16.
Patkowski.
M. (IYSO). The senslti\e perwd
Lrrrnrbrg.
for the acquisition of \yntau in a wcond
Phinney. M. (lY87).
The pro-drop parameter
(Eds. ). fcrrontercr srfrirry. Dordrecht:
Ritchie. W. (lY7S). The right-roof constraint
lrrrrguoyr trcquisirim
reworch.
in second latyuage
son. D. Lightfoot. S: Y.C. ?.lorin (Ed\.).
Ross. J.R. (1967).
In T. Roeper Y; E. Williams
in adult w:ond language scqui\ition.
New York: Acadcmlc
Iwres i,r Imlim~ s~mr.r.
acquisitwn.
Reldel.
Rizzl. L. (lY7Y). Violations of the uh-island con\trant
L. (1981).
lunpt~a_re. Lnrr,ywgr
.W. 14Y-472.
in Italian and the wbjncenc!
.\/o~trrzoi worXi/rgpcrperr
Dordrecht:
Co/~srwirrrs 01, I rzritrhlrr i/t rurrm.
In W. Ritchw (Ed.).
Srcwr</
Press.
in littgukrics
condition. In C. Dubui+
II. (Reprinted
in Rizzt.
For]\).
Doctoral dissertation. MIT.
Schachter. J. (IYSY). Testing a proposed universal. To appear m S. Gab\ & J. Schxhtrr
(Ed\.).
Littguhric
prrspecrit~es m srcorrd lmrglroge ocqlrisiriou.
Snow. C.. & Horfnagel-Hohle.
M. (IY78). The crittcal period for language acquisition: Esidence from second
language learning. Child De~dopmeur.
49.
I ICI 128.
Sportiche. D. (1081). On boundins nodes in French. L:rrguisfic
Torreeo.
Wexler.
(IY8-I).
K.. & Culicover.
White. L. (IY84a).
Rwielt~. 1. 21Y-246.
On inversion in Spanish and some oi its effects. Lirrgtrisric
P. (lY80).
Forrnol prirriplrs
ofkrrrgutrge
Irrquir~.
rrcquisi/iorr.
/.i.
I(&12Y.
Cnmbrid_ee. SIA: MIT
Pre%.
Implications of parametric variation for adult second language acquisition: An investigation
of the pro-drop parameter.
In V. Cook (Ed.).
Er;wrivrrnr~rl
opprotrches
to srcorrd lmgutrge crcyuisirimf.
Perpamon Press. Oxford.
White. L. (1981b). Universal grammar and langungr tranrfer.
University of Wisconsin-Milwaukee
Revision of a paper deli\ered
at the 13th Annual
Symposium on Current Approaches to SL.4. March 1YX-l.